Table 2.

Performance comparison of the models trained on real data versus synthetic data.

Model performances based on the “real” secondary dataset	Trained on dataset A real data (95% CI)	Trained on dataset B (synthetic data ×1) (95% CI)	Trained on dataset C (synthetic data ×2) (95% CI)	Trained on dataset D (synthetic data ×5) (95% CI)
MILO’s best models	MILO GBM	MILO SVM	MILO DNN	MILO DNN
ROC-AUC	0.95 (0.87–1)	0.83 (0.63–1)	0.91 (0.8–1)	0.55 (0.48–0.62)
Accuracy	90 (84–95)	91 (85–95)	71 (63–78)	54 (46–62)
Sensitivity	89 (83–94)	93 (87–96)	67 (59–75)	49 (40–58)
Specificity	100 (81–100)	77 (50–93)	100 (81–100)	94 (71–99)
MILO’s best RF models	MILO RF	MILO RF	MILO RF	MILO RF
ROC-AUC	0.96 (0.82–1)	0.77 (0.67–0.87)	0.87 (0.77–0.97)	0.66 (0.52–0.8)
Accuracy	89 (83–93)	71 (63–78)	74 (66–81)	56 (48–64)
Sensitivity	88 (81–93)	69 (60–76)	72 (64–80)	53 (44–61)
Specificity	100 (81–100)	88 (64–99)	88 (64–99)	82 (57–96)
Non-MILO RF models	Non-MILO RF	Non-MILO RF	Non-MILO RF	Non-MILO RF
ROC-AUC	0.97 (0.94–1)	0.73 (0.60–0.88)	0.83 (0.71–0.92)	0.68 (0.57–0.82)
Accuracy	77 (70–84)	62 (54–69)	64 (56–72)	39 (31–47)
Sensitivity	75 (66–82)	61 (52–69)	64 (55–72)	40 (32–49)
Specificity	100 (81–100)	71 (44–90)	71 (44–90)	29 (10–56)

DNN = deep neural network, GBM = gradient boosting machine, RF = random forest, SVM = support vector machine.